While multimodal recipe data is effective for alleviating data sparsity in recommender systems, existing methods primarily focus on learning common knowledge across modalities, neglecting the unique features within each one. This oversight allows user preference-irrelevant information to persist in both modality-shared and modality-specific representations, ultimately degrading recommendation performance. To address this, we propose Mirror, a novel multimodal recipe recommendation learning framework. Mirror employs a modality-disentangled representation learning method to decompose recipe features into distinct modality-shared and modality-specific components. Furthermore, it incorporates a multimodal counterfactual learning layer designed to identify and eliminate user preference-irrelevant features. This purification process refines the multimodal representations and enhances the model’s ability to capture users’ latent interests. Extensive experimental results verify the superiority of Mirror against many existing methods.